Add network syscalls

[vinc]

• Add /dev/net/tcp device file to open a TCP socket handle
• Add /dev/net/udp device file to open a UDP socket handle
• Add poll syscall
• Add connect syscall
• Add listen syscall
• Add accept syscall
• Rewrite tcp command to use the network using syscalls only
• Rewrite http command to use the network using syscalls only
• Rewrite httpd command to use the network using syscalls only
• Rewrite socket command to use the network using syscalls only
• Rewrite host command to use the network using syscalls only
• Rewrite dhcp command to use the network using syscalls only

[vinc]

Syscalls used for a TCP Client:

- OPEN /dev/net/tcp -> handle
- CONNECT handle address port
- WRITE handle buffer -> bytes written
- READ handle buffer -> bytes read
- CLOSE handle

[vinc]

We could introduce an equivalent of Unix IOCTL named CONTROL or MANAGE to replace CONNECT handle address port with MANAGE handle request response where request could be "connect 192.168.1.1 80" for example.

[vinc]

We need a poll syscall to be able to wait on multiple handles without blocking.

For example:

let list = vec![(stdin, Event::Read), (socket, Event::Read)];
loop {
  if let Some((h, e)) = syscall::poll(&list) {
    // Read from stdin or socket, whichever is ready first
  }
}

[vinc]

The updated socket command using poll allows us to have a chat interface with other computers.

The command socket -l 1234 and socket 192.168.1.100 1234 have the equivalent on Unix with nc and can be used to launch a basic chat server/client on port 1234.

What's missing is knowing when a socket connection is closed. We could add another IO::Close event to poll.

[vinc]

Poll is not blocking for now, so we could poll only one file handle at a time instead of a list. The list is only useful to reduce the number of userspace/kernelspace boundary crossing. But if poll was blocking the we would cross it only once at the beginning. And in that case we need the list because we can't poll the file handles one by one.

[vinc]

With a blocking poll the console and socket read don't need to be blocking.

[vinc]

Sometime the TCP state after a connect change directly from syn-sent to close-wait without seemingly going through established but that's only because the transition is too fast. In that case we can check if there's something that can be read to know if an error occurred or not.

[vinc]

There was some confusion with the syntax for the port number between tcp and socket on one side and http on the other side:

• tcp <host> <port>
• socket [<host>] <port>
• http <host>[:<port>][ ]<path>

We can't change http because of the path so this commit changes the first two commands to remove the confusion.

• tcp <host>:<port>
• socket [<host>:]port
• http <host>[:<port>][ ]<path>

We don't support IPv6 yet but we might want to use ! instead of : like Plan 9 does when this will be the case.

[vinc]

Finally we also introduced a special case in read to return the socket state without blocking when the given buffer is 1 byte. This is a simple hack that we are also using for reading the console where 4 bytes is read_char and something else like 256 is read_line.

[vinc]

There's two options for DHCPv4, either using the smoltcp crate's implementation of the protocol in the kernel by creating a special dhcp socket, or reimplementing the protocol in userspace using a pair of UDP sockets. The latter is the most interesting way to do it but this work could be done later in another PR.

The HTTP server will also have to be rewritten but this can be done in a separate PR too.

[vinc]

The socket file size will now show the recommended buffer size:

> list /dev/net
1446 2023-07-29 12:09:21 tcp
1458 2023-07-29 12:09:21 udp

Size: MTU - Ethernet header - IP header - TCP/UDP header

[vinc]

We can now delete a device file. This was needed to easily recreate the socket files and change their sizes without recreating the file system.